About Hy4 preview
Hy4 preview is a 770B parameter mixture-of-experts (MoE) model from Tencent, with 49B active parameters and a 1M context window. It is an open-source release under the Apache 2.0 license, positioned as the next step after Hy3 preview. The model targets long-horizon agentic work, including coding, game development, and complex document analysis.
Review
Tencent's approach with Hy4 preview centers on training the model with data from its own engineering and specialist teams. The company ran a blind evaluation with 163 internal experts across 203 engineering tasks, where Hy4 preview scored slightly ahead of GLM 5.3 and Kimi K3. The model is currently available through OpenRouter and WorkBuddy.
Key Features
- Autonomous task execution: the model runs its own tests, fixes bugs, and continues working on long-horizon tasks.
- A 770B total parameter MoE architecture that activates only 49B parameters per query for inference.
- A 1M-token context window to accommodate large codebases and lengthy documents.
- Open-source distribution under Apache 2.0, allowing modification and commercial use.
Pricing and Value
The model is listed as free, a claim that aligns with its open-source Apache 2.0 license. Access through OpenRouter may carry its own usage fees based on that platform's pricing structure. The company doesn't publish a commercial API pricing tier for Hy4 preview, so enterprise pricing remains not yet defined.
Pros
- The 1M context window handles entire code repositories or long technical documents without truncation.
- Its training on internal engineering tasks means it has seen realistic, messy work that spans hours.
- Open weight availability enables self-hosting, fine-tuning, and integration into private toolchains.
- The blind internal eval places it above two notable competitor models in engineering task completion.
Cons
- It exhibits overthinking: a user reported the model re-derived the same file layout four times before proceeding, consuming most of the context window and stalling the run.
- No long-term usage data exists. It's a preview release, and it lacks the official launch of a formal version, so stability and performance across diverse tasks are unverified.
- It's not well suited for teams needing predictable, low-latency responses for routine queries, where its tendency to verify and re-check slows output.
Hy4 preview fits developers who can tolerate a slower, more meticulous process in exchange for context-heavy agentic execution. It's best for coding, research, and document analysis tasks that stretch over long sessions, where that 1M window and iterative testing pay off. Teams running simple chat prompts or seeking rapid-fire answers will likely find the model's cautious behavior unnecessary.
Open 'Hy4 preview' Website
Your membership also unlocks:
